ACE launches self-improving AGENTS.md playbooks for code factories

• ACE launched as an open-source and hosted platform that treats AGENTS.md as a revisable operating document rather than a fixed prompt, with the author calling it "RSI for your AI coding agent's AGENTS.md" ACE launch and the repo describing "dynamic, evolving playbooks" GitHub repo.
• The implementation pitch is a code-factory workflow wired to tools like Linear, Codex, and Symphony, with the launch thread saying it was tuned with "a few billion tokens of GPT-5.4 intelligence" over the last four days launch thread.
• ACE's repo says the system records outcomes from real tasks, evolves playbooks from execution history, and aims to reduce "prompt drift" and repeated failures across coding and ops workflows repo summary.
• The security framing is explicit: the same author argues "no prompt is safe" prompt thread, citing prompt extraction as a practical reason to move critical behavior out of static hidden instructions.

ACE is now available in both open-source and hosted form, with the launch post pointing to a hosted service ACE site and a separate post linking the OSS repo GitHub repo. The core idea is to turn an agent's AGENTS.md from a static instruction file into what the repo calls an "evolving playbook" repo summary that captures successful strategies, failures, and revisions over time.

The author frames ACE as part of a larger "Code Factory" stack built around Linear, Codex, and Symphony launch thread. In the repo description, ACE is positioned as tool-agnostic and usable across coding, research, writing, analysis, and operations, with support for MCP-compatible tools including Claude Code and Codex repo summary. Self-hosted deployment is documented through Docker Compose and local development paths, while the hosted option targets teams that want the same workflow without operating the stack themselves repo summary.

ACE's launch is really a response to two operational problems: prompt drift and prompt leakage. The repo summary says it learns from "real tasks" so useful patterns are not lost between runs, and that it automatically evolves playbooks from execution history instead of relying on a single frozen prompt repo summary. That makes the system's behavior more inspectable and revisable than burying more rules inside system prompts.

The same author also argues that "no prompt is safe" prompt thread, and the attached screenshot describes a team whose internal tool exposed its full system prompt after users asked for the instructions "with some creative phrasing" [img:2|prompt extraction post]. That does not prove ACE solves prompt extraction outright, but it does explain the design choice: move durable operating logic into versionable playbooks and execution traces rather than treating hidden prompts as a reliable security boundary.